\documentclass[12pt]{article}
\usepackage{mypackage}


\begin{document}

\title{Can you spot a scam? \\ Measuring and improving scam identification ability}

\author{Elif Kubilay$^1$ \textcircled{r} Eva Raiber$^2$ \textcircled{r} Lisa Spantig$^{3,1}$ \textcircled{r} Jana Cahlíková$^4$ \textcircled{r} Lucy Kaaria$^5$}
\date{%
	$^1${\small University of Essex},%
	$^2${\small Aix-Marseille Universit\'e, CNRS, AMSE}\\%
	$^3${\small RWTH Aachen University},%
	$^4${\small University of Bonn},%
	$^5${\small University of Nairobi and HOPAWI}\\%
}
\maketitle
\begin{abstract} 
The expansion of digital financial services leads to severe consumer protection issues such as fraud and scams. As these potentially decrease trust in digital services, especially in developing countries, avoiding victimization has become an important policy objective. In an online experiment, we first investigate how well individuals in Kenya identify phone scams using a novel measure of scam identification ability. We then test the effectiveness of scam education, a commonly used approach by organizations for fraud prevention. We find that common tips on how to spot scams do not significantly improve individuals' scam identification ability, i.e., the distinction between scams and genuine messages. This null effect is driven by an increase in correctly identified scams and a decrease in correctly identified genuine messages, indicating overcaution. Additionally, we find suggestive evidence that genuine messages with scam-like features are misclassified more often, highlighting the importance of a careful design of official communication.


\noindent 
\text{JEL Codes: D14, D18, G53, O12} \\
\text{Keywords: Consumer protection, consumer fraud, digital financial services, scam} \\ 
\text{susceptibility, scam education, Kenya}
\end{abstract}

\newpage 

%%%%%%%%%%%%%%%%%%%%%%%%%
%%% Figure 1

\begin{figure}[H]
    \centering
    \includegraphics[scale=0.2]{3_figures/tips_all.png}
    \caption{Tips treatment}
    \label{fig:info_treatment}
    \begin{minipage}[t]{0.85\linewidth} \setstretch{0.8}
                { \vspace{0.1cm}\footnotesize{\textit{Notes:} Tips treatment was designed based on commonly communicated tips in Kenya. The graphic was ``animated,'' such that the pieces of information would be shown step-by-step. Participants clicked through this animation at their own speed, i.e., they hit the ``continue" button five times before they see the overall graphic.}}
            \end{minipage}
\end{figure}

\newpage
%%%%%%%%%%%%%%%%%%%%%%%%%
%%% Table 1

 
\begin{table}[H]\caption{Correlates of Scam Identification Ability and Confidence}\label{tab:correlates_SIA}
	\begin{center}
		\renewcommand{\arraystretch}{1}
		\resizebox{15cm}{!}{
			\input{4_tables/table1}} \\
			\begin{minipage}[t]{0.85\linewidth} \setstretch{0.8}
                { \vspace{0.1cm}\footnotesize{\textit{Notes:} Dependent variables are the share of correctly identified messages (SIA) in block 1 and average confidence ratings in block 1. \textit{Female, Post-Secondary Education, Formal Employment, Low Trust in DFS, Contacted less than 1 week ago}, and \textit{victim of a scammer} are binary indicators, \textit{Low Income} and \textit{High use of different DFS} are binary indicators for median splits. All variables rely on self-reports. All specifications control for the order of the two blocks and failing the attention check.  The displayed coefficients are from OLS regressions. Robust standard errors are in parenthesis. Asterisks indicate that the estimate is statistically significant at the 1\% $^{***}$, 5\% $^{**}$, and 10\% $^{*}$ levels.}}
            \end{minipage}
	\end{center}
\end{table}

\newpage

%%%%%%%%%%%%%%%%%%%%%%%%%
%%% Table 2

\begin{table}[H]
     \caption{Treatment Effects}\label{tab:tr_effects}
     \begin{subtable}[t]{\linewidth}
     		\renewcommand{\arraystretch}{1}
    \begin{center}
     		\caption*{\textit{Panel 1: Main Outcomes}}
			\resizebox{15cm}{!}{
				\input{4_tables/table2_panel1}} 
		\end{center}
     \vspace{-1.8em}
			\begin{center} 
			\caption*{\textit{Panel 2: Secondary Outcomes}}
			\resizebox{15cm}{!}{
				\input{4_tables/table2_panel2}}
		\begin{minipage}[t]{0.85\linewidth} \setstretch{0.8}
{ \vspace{0.1cm}\footnotesize{\textit{Notes:}  In Panel 1, the dependent variables are the share of correctly identified messages(SIA) in block 2, the share of correctly identified scams in block 2, the share of correctly identified non-scams in block 2, and the average confidence ratings in block 2 for all messages (confidence in SIA), for the scam messages, and for the non-scam messages. In Panel 2, the dependent variables are a binary indicator for low trust in DFS, the time spent on SIA in block 2, a binary indicator for classifying all scams correctly in block 2, and a binary indicator for classifying all non-scams correctly in block 2. All specifications include an indicator for the incentives treatment, the value of the outcome variable in block 1 (except for trust, which was only measured after block 2), and the full set of controls, i.e., variables displayed in Table~\ref{tab:correlates_SIA} (female, age, post-secondary education, low income, formal employment, low trust in DFS (except for the effect on trust), above median use of different DFS, contacted less than one week ago, victim of a scammer), as well as indicators for the order of the two blocks and failing the attention check. $Tips^{U}$ and $Tips^{I}$ refer to Tips (unincentivized) and Tips (incentivized), respectively. The displayed coefficients are from OLS regressions.  Robust standard errors are in parenthesis. Asterisks indicate that the estimate is statistically significant at the 1\% $^{***}$, 5\% $^{**}$, and 10\% $^{*}$ levels. }}%
\end{minipage}
	\end{center}
    \end{subtable}
 \end{table}

\newpage
%%%%%%%%%%%%%%%%%%%%%%%%%
%%% Figure 2


\begin{figure}[H]
    \centering
    \caption{Treatment Effect Heterogeneity} 
    \begin{subfigure}{1\textwidth}
    \centering
        \caption{Scam Identification Ability (SIA)}\label{fig:heterogeneity_SIA}
        \includegraphics[scale=0.4]{3_figures/figure2_panel1.png}
    \end{subfigure}
    \begin{subfigure}{1\textwidth}
    \centering
        \caption{Confidence}\label{fig:heterogeneity_confidence}
       \includegraphics[scale=0.4]{3_figures/figure2_panel2.png}
    \end{subfigure}
     \begin{minipage}[t]{0.85\linewidth} \setstretch{0.8}
                { \vspace{0.1cm}\footnotesize{\textit{Notes:} Figures plot the OLS coefficients and the 90\% and 95\% confidence intervals from the estimating regressions in Panel 1, Table~\ref{tab:tr_effects} (Column 1 for SIA and Column 4 for Confidence) separately for the different subcategories. }}
            \end{minipage}
\end{figure}

\newpage
%%%%%%%%%%%%%%%%%%%%%%%%%
%%% Figure 3

\begin{figure}[H]
    \centering
    \includegraphics[scale=0.25]{3_figures/figure3}
    \caption{Vignette-level effects by whether the message contains a scam marker}
    \label{fig:vigmarker}
     \begin{minipage}[t]{0.85\linewidth} \setstretch{0.8}
                { \vspace{0.1cm}\footnotesize{\textit{Notes:} Figures plot the average marginal effects of triple-differences estimation with 95\% confidence intervals based on standard errors clustered at the respondent level (see also Appendix~\ref{sec:app_vignettes}). Scam Marker is an indicator for whether the message contains at least one of the scam markers the tips warn about. The left panel contains all vignettes, the center panel focuses on scams, and the right panel on non scams. For ease of exposition, only the control and the Tips (unincentivized) treatment are displayed. The empirical specification contains the full set of interactions and demographic controls.

}}
            \end{minipage}
\end{figure}



\newpage







\section*{Online Appendix}
\renewcommand{\thesubsection}{\Alph{subsection}}
\setcounter{table}{0}
\setcounter{figure}{0}


\subsection{Additional Tables}
\pagenumbering{arabic}
\renewcommand{\thetable}{A\arabic{table}}
\renewcommand{\thefigure}{A\arabic{figure}}

%%%%%%%%%%%%%%%%%%%%%%%%%
%%% Table A1

\begin{table}[H]\caption{Overview of vignettes}\label{tab:vignettes_overview}
    \centering
    \resizebox{14cm}{!}{
        \input{4_tables/table_vignettes_senders.tex}
    }
    \begin{minipage}[t]{0.85\linewidth} \setstretch{0.8}
                { \vspace{0.1cm}\footnotesize{\textit{Notes:} Vignettes kept the original wording of the screenshots and were visually harmonized, i.e., all vignettes were displayed on the same phone with the same signal strength, battery level, etc. (see also Figure~\ref{fig:vignette_example}). The order of the blocks was randomized at the individual level, as was the order of vignettes within a block. As blocks might be different for various reasons, we always control for the order of the blocks in the analysis.}}
            \end{minipage}
\end{table}
\newpage

%%%%%%%%%%%%%%%%%%%%%%%%%
%%% Table A2

\begin{landscape}
\begin{table}[H]\caption{Balance}\label{tab:balance}
	\begin{center}
		\renewcommand{\arraystretch}{1}
		\resizebox{24cm}{!}{
			\input{4_tables/tableA2}} \\
			\begin{minipage}[t]{\linewidth} \setstretch{0.8}
                { \vspace{0.1cm}\footnotesize{\textit{Notes:} Asterisks indicate that the difference is statistically significant at the 1\% $^{***}$, 5\% $^{**}$, and 10\% $^{*}$ levels.}}
            \end{minipage}
	\end{center}
\end{table}
\end{landscape}
\newpage
%%%%%%%%%%%%%%%%%%%%%%%%%
%%% Table A3
\begin{table}[H] \caption{Sample and Kenyan Population}\label{tab:sample_comparison}
    \centering
    \renewcommand{\arraystretch}{1}
\resizebox{11cm}{!}{
    \input{4_tables/table_comparison_v2}} 
\end{table}
\newpage
%%%%%%%%%%%%%%%%%%%%%%%%%
%%% Table A4

\begin{table}[H]\caption{Sample Characteristics}\label{tab:summary_stats}
	\begin{center}
		\renewcommand{\arraystretch}{1.2}
		\resizebox{14cm}{!}{
			\input{4_tables/tableA4}} \\
			\begin{minipage}[t]{0.62\linewidth} \setstretch{0.8}
                { \vspace{0.1cm}\footnotesize{}}
            \end{minipage}
	\end{center}
\end{table}

\newpage
%%%%%%%%%%%%%%%%%%%%%%%%%
%%% Table A5

\begin{table}[H]
     \caption{The Effects of Incentives}\label{tab:incentive_effects}
     \begin{subtable}[t]{\linewidth}
     		\renewcommand{\arraystretch}{1}
     		\begin{center}
     		\caption*{\textit{Panel 1: Outcomes in Block 1}}
			\resizebox{15cm}{!}{
				\input{4_tables/tableA5_panel1}} 
		\end{center}
     \vspace{-1.8em}
			\begin{center} 
			\caption*{\textit{Panel 2: Outcomes in Block 2}}
			\resizebox{15cm}{!}{
				\input{4_tables/tableA5_panel2}}
		\begin{minipage}[t]{0.85\linewidth} \setstretch{0.8}
{ \vspace{0.1cm}\footnotesize{\textit{Notes:} In Panels 1 and 2, dependent variables are the share of correctly identified messages (SIA), the share of correctly identified scams, the share of correctly identified non-scams, and the average confidence ratings in block 1 and 2, respectively. All specifications include indicators for the tips treatments, and the full set of controls, i.e., variables displayed in Table~\ref{tab:correlates_SIA} (female, age, post-secondary education, low income, formal employment, low trust in DFS, above median use of different DFS, contacted less than one week ago, victim of a scammer), as well as indicators for the order of the two blocks and failing the attention check. The displayed coefficients are from OLS regressions. Robust standard errors are in parenthesis. Asterisks indicate that the estimate is statistically significant at the 1\% $^{***}$, 5\% $^{**}$, and 10\% $^{*}$ levels.}}%
\end{minipage}
	\end{center}
    \end{subtable}
 \end{table}

\newpage

%%%%%%%%%%%%%%%%%%%%%%%%%
%%% Table A6

\begin{table}[H]\caption{Confidence weighted by SIA}\label{tab:explore_confidence_reverse}
	\begin{center}
		\renewcommand{\arraystretch}{1}
		\resizebox{15cm}{!}{
			\input{4_tables/tableA6}} \\
			\begin{minipage}[t]{0.85\linewidth} \setstretch{0.8}
                { \vspace{0.1cm}\footnotesize{\textit{Notes:} The dependent variables are weighted confidence in block 2 for all messages, scams, and non-scams, respectively. Weighted confidence ranges from -1 to 1, where 1 means fully confident and perfect SIA, whereas -1 means fully confident and no correctly classified vignette. 
                All specifications include an indicator for the incentives treatment, the value of the outcome variable in block 1, and the full set of controls, i.e., variables displayed in Table~\ref{tab:correlates_SIA} (female, age, post-secondary education, low income, formal employment, low trust in DFS, above median use of different DFS, contacted less than one week ago, victim of a scammer), as well as indicators for the order of the two blocks and failing the attention check. $Tips^{U}$ and $Tips^{I}$ refer to Tips (unincentivized) and Tips (incentivized), respectively. The displayed coefficients are from OLS regressions. Robust standard errors are in parenthesis. Asterisks indicate that the estimate is statistically significant at the 1\% $^{***}$, 5\% $^{**}$, and 10\% $^{*}$ levels.}}
            \end{minipage}
	\end{center}
\end{table}

\newpage

\subsection*{Additional Figures}
%%%%%%%%%%%%%%%%%%%%%%%%%
%%% Figure A1
\begin{figure}[H]
    \centering
    \includegraphics[scale=0.5]{3_figures/A_M1.png}
    \caption{Example Vignette}
    \label{fig:vignette_example}
\end{figure}

\newpage

%%%%%%%%%%%%%%%%%%%%%%%%%
%%% Figure A2
\begin{figure}[H]
    \centering
    \caption{SIA and Confidence in Block 1} \label{fig:siaconf_hist}
    \begin{subfigure}{0.75\textwidth}
    \centering
        \caption{Distribution of scam identification ability (SIA)}
        \includegraphics[scale=0.4]{3_figures/figureA2_panel1.png}
    \end{subfigure}
    \begin{subfigure}{0.75\textwidth}
    \centering
        \caption{Distribution of Confidence in SIA}
       \includegraphics[scale=0.4]{3_figures/figureA2_panel2.png}
    \end{subfigure}
\end{figure}



\clearpage

\subsection{Robustness}\label{sec:app_robust}
\setcounter{table}{0}
\renewcommand{\thetable}{B\arabic{table}}


%%%%%%%%%%%%%%%%%%%%%%%%%
%%% Table B1

\begin{table}[H]\caption{Attention check and main treatment effects}\label{tab:attention}
	\begin{center}
		\renewcommand{\arraystretch}{1}
		\resizebox{15cm}{!}{
			\input{4_tables/tableB1}} \\
			\begin{minipage}[t]{0.85\linewidth} \setstretch{0.8}
                { \vspace{0.1cm}\footnotesize{\textit{Notes:} Columns 1-6 include only those who passed the attention check. Column 7 includes all participants. Dependent variables are the share of correctly identified messages (SIA) in block 2, the share of correctly identified scams in block 2, the share of correctly identified non-scams in block 2, the average confidence ratings in block 2 for all messages (confidence in SIA), for scam messages, and for non-scam messages and an indicator for failing the attention check. All specifications include an indicator for the incentives treatment, the value of the outcome variable in block 1 (except for the attention check which was only administered once), and the full set of controls, i.e., variables displayed in Table~\ref{tab:correlates_SIA} (female, age, post-secondary education, low income, formal employment, low trust in DFS, above median use of different DFS, contacted less than one week ago, victim of a scammer), as well as indicators for the order of the two blocks and failing the attention check (except for the last specification where the attention check is the dependent variable). $Tips^{U}$ and $Tips^{I}$ refer to Tips (unincentivized) and Tips (incentivized), respectively. The displayed coefficients are from OLS regressions. Robust standard errors are in parenthesis. Asterisks indicate that the estimate is statistically significant at the 1\% $^{***}$, 5\% $^{**}$, and 10\% $^{*}$ levels.}}
            \end{minipage}
	\end{center}
\end{table}

\newpage
%%%%%%%%%%%%%%%%%%%%%%%%%
%%% Table B2

 \begin{table}[H]
     \caption{Treatment Effects: Additional Robustness Checks}\label{tab:main_robustness}
     \begin{subtable}[t]{\linewidth}
     		\renewcommand{\arraystretch}{1}
     		\begin{center}
     		\caption*{\textit{Panel 1: Excluding the Baseline SIA and Confidence}}
			\resizebox{10cm}{!}{
				\input{4_tables/tableB2_panel1}} 
		\end{center}
     \vspace{-1.8em}
			\begin{center} 
			\caption*{\textit{Panel 2: Incentivized Sample}}
			\resizebox{10cm}{!}{
				\input{4_tables/tableB2_panel2}}
       \end{center}
     \vspace{-1.8em}
			\begin{center} 
			\caption*{\textit{Panel 3: Unincentivized Sample}}

			\resizebox{10cm}{!}{
				\input{4_tables/tableB2_panel3}}
		\begin{minipage}[t]{0.58\linewidth} \setstretch{0.8}
{ \vspace{0.1cm}\footnotesize{\textit{Notes:}  The dependent variables are the scam identification ability share of correctly identified messages (SIA) in block 2, the share of correctly identified scams in block 2, the share of correctly identified non-scams in block 2, and the average confidence ratings in block 2 for all messages (confidence in SIA), for the scam messages, and for the non-scam messages. In Panel 1, all specifications include an indicator for the incentives treatment. In Panels 2 and 3, the sample is restricted to incentivized and unincentivized respondents, respectively. All specifications include the full set of controls, i.e., variables displayed in Table 1 (female, age, post-secondary education, low income, formal employment, low trust in DFS, above median use of different DFS, contacted less than one week ago, victim of a scammer), as well as indicators for the order of the two blocks and failing the attention check. $Tips^{U}$ and $Tips^{I}$ refer to Tips (unincentivized) and Tips (incentivized), respectively. Displayed coefficients are from OLS regressions. Robust standard errors are in parenthesis. Asterisks indicate that the estimate is statistically significant at the 1\% $^{***}$, 5\% $^{**}$, and 10\% $^{*}$ levels. }}%
\end{minipage}
	\end{center}
    \end{subtable}
 \end{table}

\newpage
%%%%%%%%%%%%%%%%%%%%%%%%%
%%% Table B2

\begin{table}[H]
     \caption{Treatment Effects: Varying Control Variables}\label{tab:main_controls}
     \begin{subtable}[t]{\linewidth}
     		\renewcommand{\arraystretch}{1}
     		\begin{center}
     		\caption*{\textit{Panel 1: Scam Identification Ability}}
			\resizebox{10cm}{!}{
				\input{4_tables/tableB3_panel1}} 
		\end{center}
     \vspace{-1.8em}
			\begin{center} 
			\caption*{\textit{Panel 2: Confidence}}
			\resizebox{10cm}{!}{
				\input{4_tables/tableB3_panel2}}
		\begin{minipage}[t]{0.58\linewidth} \setstretch{0.8}
{ \vspace{0.1cm}\footnotesize{\textit{Notes:}  Dependent variables in Panel 1 and Panel 2 are the share of correctly identified messages (SIA) and average confidence in block 2, respectively. Demographic controls include gender, age, post-secondary education, low income, formal employment), design controls include indicators for order of blocks and a dummy for attention check, controls for scam experience include low trust in DFS, above median use of different DFS, contacted less than one week ago, and victim of a scammer. All specifications include an indicator for the incentives treatment and the baseline value of the outcome variable. $Tips^{U}$ and $Tips^{I}$ refer to Tips (unincentivized) and Tips (incentivized), respectively. Displayed coefficients are from OLS regressions. Robust standard errors are in parenthesis. Asterisks indicate that the estimate is statistically significant at the 1\% $^{***}$, 5\% $^{**}$, and 10\% $^{*}$ levels. }}%
\end{minipage}
	\end{center}
    \end{subtable}
 \end{table}





\subsection{Scam Data Collection and SIA Measurement} \label{sec:app_datacollectsia}

%%%%%%%%%%%%%%%%%%%%%%%%%
%%% Table C1/D1

\subsubsection{Scam-related Keywords}\label{sec:app_keywords}
\setcounter{table}{0}
\renewcommand{\thetable}{C\arabic{table}}
\begin{table}[H]\caption{Scam Keywords}\label{tab:keywords}
	\begin{center}
		\renewcommand{\arraystretch}{1}
		\resizebox{15cm}{!}{
			\input{4_tables/table_keywords.tex}} \\
			\begin{minipage}[t]{0.85\linewidth} \setstretch{0.8}
                { \vspace{0.1cm}\footnotesize{\textit{Notes:} Keywords were determined based on discussions with Kenyan DFS experts.}}
            \end{minipage}
	\end{center}
\end{table}

\clearpage


\subsection{Vignette-Level Analysis}\label{sec:app_vignettes}
\setcounter{figure}{0}
\renewcommand{\thefigure}{D\arabic{figure}}
\renewcommand{\theequation}{D\arabic{equation}}

%%%%%%%%%%%%%%%%%%%%%%%%%
%%% Figure D1
\begin{figure}[H]
    \centering
    \includegraphics[scale=0.25]{3_figures/figureD1.png}
    \caption{Vignette-level effects by whether the message contains a scam marker}
    \label{fig:vigmarker_incentivized}
     \begin{minipage}[t]{0.85\linewidth} \setstretch{0.8}
                { \vspace{0.1cm}\footnotesize{\textit{Notes:} Figures plot the average marginal effects from triple-differences estimation with 95\% confidence intervals based on standard errors clustered at the respondent level (see also Appendix~\ref{sec:app_vignettes}). Scam Marker is an indicator for whether the message contains at least one of the scam markers the tips warn about. The left panel contains all vignettes, the center panel focuses on scams, and the right panel on non scams. For ease of exposition, only the control and the Tips (incentivized) treatment are displayed. The econometric specification contains the full set of interactions and demographic controls.}}
            \end{minipage}
\end{figure}
\newpage

%%%%%%%%%%%%%%%%%%%%%%%%%
%%% Figure D2

\begin{figure}[H]
    \centering
    \includegraphics[scale=0.25]{3_figures/figureD2.png}
    \caption{Vignette-level effects by difficulty of the vignette
}
    \label{fig:vigdif}
    \begin{minipage}[t]{0.85\linewidth} \setstretch{0.8}
                { \vspace{0.1cm}\footnotesize{\textit{Notes:} Figures plot the average marginal effects from triple-differences estimation with 95\% confidence intervals based on standard errors clustered at the respondent level (see also Appendix~\ref{sec:app_vignettes}). Vignettes are coded as difficult if they are misclassified more often than the median vignette in block 1.  The left panel contains all vignettes, the center panel focuses on scams, and the right panel on non scams. For ease of exposition, only the control and the Tips (unincentivized) treatment are displayed. The econometric specification contains the full set of interactions and demographic controls.}}
            \end{minipage}
\end{figure}
\clearpage





\end{document}



